Disentangling Shared and Private Neural Dynamics with SPIRE: A Latent Modeling Framework for Deep Brain Stimulation

TL;DR

SPIRE introduces a nonlinear, shared–private latent framework for multi-region neural data under stimulation. By training on baseline activity and using cross-regional alignment plus disentanglement losses, SPIRE recovers cross-regional structure and reveals how DBS reorganizes network dynamics, with shared latents encoding stimulation-specific signatures that generalize across sites and frequencies. On synthetic benchmarks, SPIRE outperforms linear and some nonlinear baselines; on pediatric DBS data, it demonstrates explicit disentanglement of shared vs. private dynamics and frequency-dependent reorganization of BG–thalamo–cortical coordination. The approach provides a practical, reproducible tool for analyzing perturbation-evoked multi-region neural dynamics and offers neuroscientific insights into how stimulation modulates cross-regional coordination at the network level.

Abstract

Disentangling shared network-level dynamics from region-specific activity is a central challenge in modeling multi-region neural data. We introduce SPIRE (Shared-Private Inter-Regional Encoder), a deep multi-encoder autoencoder that factorizes recordings into shared and private latent subspaces with novel alignment and disentanglement losses. Trained solely on baseline data, SPIRE robustly recovers cross-regional structure and reveals how external perturbations reorganize it. On synthetic benchmarks with ground-truth latents, SPIRE outperforms classical probabilistic models under nonlinear distortions and temporal misalignments. Applied to intracranial deep brain stimulation (DBS) recordings, SPIRE shows that shared latents reliably encode stimulation-specific signatures that generalize across sites and frequencies. These results establish SPIRE as a practical, reproducible tool for analyzing multi-region neural dynamics under stimulation.

Figures (12)

• Figure 1: Schematic of model with main goals
• Figure 2: Visual comparison of shared latents of region 1 in the nonlinear regime (D1). SPIRE (left panels) and DLAG (right panels) are CCA-aligned to the ground truth. Solid lines show trial-averaged latent trajectories, with shaded areas denoting $\pm$ SEM across trials. SPIRE achieves consistently higher alignment with ground-truth shared latents. Both models were evaluated on the identical dataset; apparent GT differences reflect that CCA alignment was performed separately for each model.
• Figure 3: Statistical comparison of DLAG and SPIRE. Mean $\pm$ SD CCA across four seeds, faceted by dataset regime. SPIRE outperforms DLAG in nonlinear (D1) and time-varying delay (D2) regimes for shared latents, and is statistically significantly better in retrieving private latents. Significance: $*:p<0.05$, $**:p<0.01$, $***:p<0.001$.
• Figure 4: Shared–private structure differs across subjects. Panels (a,b) show 3D UMAPs: in S3_R, shared GPi/STN embeddings largely overlap while shared and private clusters of both regions remain distinct; in S8_R, private vs shared of both regions separate strongly and shared embeddings intermix less. Panels (c,d) plot paired latent traces for three dims: S3_R exhibits well-aligned shared trajectories (small lag) with clearly distinct private trajectories. In S8_R, shared GPi/STN trajectories are still phase-aligned but have very low amplitude and primarily reflect slow baseline co-modulation, whereas private trajectories carry larger, region-specific dynamics that dominate the variance. These examples complement the cohort-level variance partition (Fig. \ref{['fig:var_stack']}).
• Figure 5: Validation of shared–private separation. (a) CCA between latent subspaces shows that shared GPi and STN latents are highly correlated with each other, while their leakage into private latents is weak—supporting effective disentanglement. (b) Reconstruction accuracy on held-out test data indicates that full latents (shared+private) yield the lowest error; private-only performs worst; while shared latents enable substantially better reconstructions, with same-region shared outperforming cross-region shared. Together, these analyses confirm that shared latents encode the dominant cross-regional structure, while private latents capture residual, region-specific variance. Boxplots: median, IQR, whiskers $1.5\times$IQR. Statistics: linear mixed-effects with Tukey correction ($n=17$) tukey1949comparing. Significance: $*:p<0.05$, $**:p<0.01$, $***:p<0.001$.